Breaking News
Chinese AI startup DeepSeek has introduced DeepSeek-V3.1, the latest upgrade to its flagship model. The update emphasizes faster performance, a hybrid reasoning architecture, and compatibility with China-made chips, signaling the company's push to align with the country's emerging semiconductor ecosystem amidst U.S. export restrictions.
DeepSeek said the upgraded model includes a hybrid inference design that allows it to switch between reasoning and non-reasoning modes. The new “deep thinking” toggle is available via its app and web platform, making the capability accessible to users.
In a post on WeChat, DeepSeek highlighted that the UE8M0 FP8 precision format in V3.1 is optimized for “soon-to-be-released next-generation domestic chips,” though it did not name specific hardware vendors. FP8, an 8-bit floating-point format, reduces memory use and speeds up processing compared to traditional methods. Separately, DeepSeek will revise pricing for its API—used by developers to integrate the model into applications—beginning September 6.
This upgrade comes amid broader technical hurdles. DeepSeek has struggled to train its newer R2 model on Huawei’s Ascend chips, facing delays and ultimately reverting to Nvidia hardware for model training while relegating Ascend chips to inference tasks only. The incident underscores the limitations Chinese-made hardware still faces.
Still, DeepSeek’s series of upgrades highlight its ongoing competition with global AI rivals. Earlier releases — including V3-0324 in March and an R1 update in May — have helped it position itself strongly against Western models like OpenAI’s, especially by offering advanced capabilities at lower operational cost. Experts say these efforts reflect China’s drive for AI self-reliance and the push to integrate domestic chips within key technology stacks — even as the path remains uneven.
DeepSeek said the upgraded model includes a hybrid inference design that allows it to switch between reasoning and non-reasoning modes. The new “deep thinking” toggle is available via its app and web platform, making the capability accessible to users.
In a post on WeChat, DeepSeek highlighted that the UE8M0 FP8 precision format in V3.1 is optimized for “soon-to-be-released next-generation domestic chips,” though it did not name specific hardware vendors. FP8, an 8-bit floating-point format, reduces memory use and speeds up processing compared to traditional methods. Separately, DeepSeek will revise pricing for its API—used by developers to integrate the model into applications—beginning September 6.
This upgrade comes amid broader technical hurdles. DeepSeek has struggled to train its newer R2 model on Huawei’s Ascend chips, facing delays and ultimately reverting to Nvidia hardware for model training while relegating Ascend chips to inference tasks only. The incident underscores the limitations Chinese-made hardware still faces.
Still, DeepSeek’s series of upgrades highlight its ongoing competition with global AI rivals. Earlier releases — including V3-0324 in March and an R1 update in May — have helped it position itself strongly against Western models like OpenAI’s, especially by offering advanced capabilities at lower operational cost. Experts say these efforts reflect China’s drive for AI self-reliance and the push to integrate domestic chips within key technology stacks — even as the path remains uneven.
See What’s Next in Tech With the Fast Forward Newsletter
Tweets From @varindiamag
Nothing to see here - yet
When they Tweet, their Tweets will show up here.
